Automatic Neural Network-Based Seismic Facies Classification Using Pseudo-Labels

In this paper we propose a method of seismic facies labeling. Seismic facies labeling task consists of assigning specific geological rock types to the pixels in the seismic cube. In our research we use open-source fully annotated 3D geological model of the Netherlands F3 Block. The dataset is divided into training and test cubes. We use the former to train a state-of-the-art deep learning neural network, adding a 3D conditional random field (CRF) layer as a postprocessing step. We apply the pseudo-labeling technique, where the labels of the test dataset are predicted and added to the training set to get more accurate final prediction. To diversify the training dataset, we also apply different types of augmentations, including a domain specific image warping technique. Using the trained network, we predict the facies labels on the test dataset and compute various metrics. The results suggest superior network performance over the existing baseline model.